Tutorials

Six self-contained courses. Start with the plain-language guide, then go as deep as you like with the five technical courses — no fluff, just the maths and the mechanics.

Inside the Machines

A plain-language tour of how modern AI models actually work — LLMs, encoders, encoder–decoders, mixture-of-experts, and diffusion — told as one continuous story with the mathematics removed and the intuition left standing. No calculus, linear algebra, or probability assumed. The gentle on-ramp to the five technical courses below.

How LLMs Work

A 10-lesson course tracing the full lifecycle of a large language model: from raw text tokens through self-attention and transformer blocks, all the way to training, fine-tuning with RLHF, and inference-time behaviour. Assumes comfort with linear algebra and calculus.

How Diffusion Models Work

A 10-lesson course on the mathematics of diffusion models: the forward noising process, the variational lower bound, the simplified training objective, the U-Net backbone, and fast samplers like DDIM. More probability-heavy than the LLM course; picks up where it leaves off.

How Encoder LLMs Work

A 10-lesson course on the mechanics of encoder-style models — the BERT family. Covers bidirectional self-attention, masked language modeling, how representations become predictions, pretraining dynamics, and fine-tuning. Ends with a practical landscape of RoBERTa, ALBERT, ELECTRA, DeBERTa, and ModernBERT.

How Mixture-of-Experts LLMs Work

A 10-lesson course on the mechanics of Mixture-of-Experts language models — the Switch / Mixtral / DeepSeek-MoE family. Covers the router and gating network, sparse top-k computation, load balancing, expert parallelism, training dynamics, and inference. Assumes comfort with dense decoder-only LLMs.

How Encoder–Decoder LLMs Work

A 10-lesson course on sequence-to-sequence transformers — the T5 / BART family. Covers the bidirectional encoder, the causal decoder, cross-attention as the one new mechanism, the prefix-LM, denoising pretraining objectives (span corruption, BART), teacher forcing, decoding against a frozen memory, and fine-tuning with distillation. Assumes comfort with decoder-only LLMs.